Multimodal Ground Based SAR Raw Data and Optical Images for Object Classification
Description
The dataset consists of raw Ground-Based Synthetic Aperture Radar (GBSAR) measurements and optical RGB images acquired from controlled indoor scenes containing one or two objects. Radar data were recorded using the custom-developed GBSAR-Pi system operating at 24 GHz, while optical images provide a visual reference for each unique scene configuration. GBSAR-Pi operates in stop-and-go mode using a 24 GHz Innosent IVS-362 FMCW radar module as sensor. The sensor was positioned 40 cm from the observed scene and translated along a 60 cm synthetic aperture with a step size of 1 cm. At each position, 1024 frequency samples were acquired for both the in-phase (I) and quadrature (Q) channels. Ten consecutive FMCW chirps were averaged independently for the I and Q channels to improve the signal-to-noise ratio. Objects The dataset contains measurements of seven physical objects with different dimensions: - three metallic objects (A1, A2), - three glass objects (G1, G2, G3), - two plastic objects (P1, P2). Each scene contains either one object or two objects placed at predefined locations inside the measurement area. Scene layout The measurement area contains six predefined object positions. Positions are numbered from left to right. Positions 1, 2, and 3 are farther from the radar sensor, while positions 4, 5, and 6 are closer to the radar. Radar data (.txt format) Each radar acquisition consists of two matrices representing the in-phase (I) and quadrature (Q) components of the intermediate-frequency signal. Matrix dimensions are 1024 × 60, where 1024 corresponds to the number of frequency samples within one FMCW chirp and 60 represents the number of synthetic aperture positions. The availability of both I and Q channels enables users to reconstruct complex-valued radar signals for SAR image reconstruction and complex-valued machine learning methods. Optical images (.jpg format) Optical RGB images were acquired using a fixed camera positioned 50 cm from the scene. The optical images correspond to unique scene configurations rather than individual radar acquisitions and provide visual information about object identity, material, and spatial arrangement. Radar filenames follow the convention [position][object]_[position][object]_[bandwidth]_[step size]_[polarization]_[scan direction]_[acquisition date]_[acquisition ID]_[signal component].txt Examples: 6P2_1300MHz_1.0cm_hh_R_23012026_16_Q 1A1_6P2_1300MHz_1.0cm_hh_R_23012026_16_I Dataset structure /optical_images /radar_data The dataset contains 712 radar measurements and 168 optical images. This research was supported by the NextGenerationEU framework through the project "DEEPWAVE" at the University of Zagreb Faculty of Electrical Engineering and Computing.
Files
Steps to reproduce
The dataset was acquired using the custom-developed GBSAR-Pi system. The platform consists of a Raspberry Pi 4B controlling an Innosent IVS-362 24 GHz FMCW radar module mounted on a motorized linear rail. A digital-to-analog converter generates the FMCW chirp by controlling the radar module's voltage-controlled oscillator, while an analog-to-digital converter acquires the intermediate-frequency I and Q signals. During acquisition, the radar platform operates in stop-and-go mode. At each aperture position, 1024 frequency samples are acquired for both I and Q channels. Ten consecutive chirps are averaged independently before storage. After each acquisition, the platform is translated by 1 cm using a stepper motor until the complete 60 cm synthetic aperture is scanned. Objects are positioned at predefined locations 40 cm from the radar sensor. Optical RGB images are acquired separately using a fixed camera positioned 50 cm from the scene. Both radar and camera positions remain unchanged throughout the acquisition campaign. For measurements acquired with scan direction "L", the aperture dimension should be reversed (e.g., using numpy.flip(data, axis=1)) before SAR image reconstruction because the radar platform moved in the opposite direction. The I and Q matrices can be combined into a complex-valued signal according to S = I + jQ for further radar signal processing or SAR image reconstruction.
Institutions
- University of ZagrebZagreb, Zagreb